The Next AI Filing Failure Will Be a Supervision Failure

Most coverage of AI in legal practice still sounds like a sanctions story. A bad filing appears, a judge gets angry, a fine is issued, and everyone says the obvious lesson is "check your citations." That lesson is true, but it is not deep enough. The more useful lesson is structural: if fabricated authority reaches a signed filing, the underlying failure is usually supervision, not merely software output.

The American Bar Association's June 1, 2026 article on ethics and generative AI makes that point more clearly than many court headlines do. The article describes recent hallucination incidents and then returns to the professional obligations that were already on the books: competence, confidentiality, communication, and review. In other words, the ABA is not treating AI hallucination as a weird new exception to legal ethics. It is treating it as another setting where ordinary professional duties still govern.

Why This Is Bigger Than Rule 11 Alone

Rule 11 matters. So does Rule 3.3, the duty of candor to the tribunal. But if the conversation stops there, lawyers may treat AI review as a last-minute brief-cleaning step. That is too late. The supervision duty begins earlier, when a lawyer decides how AI research will be used, who will check it, what sources must be opened, and what content may never move forward without human verification.

That is a supervision framework, not a prompt-engineering framework. Model Rule 1.1 asks whether the lawyer is competent to use the tool. Model Rule 1.6 asks whether client information is being exposed. Model Rules 5.1 and 5.3 ask whether subordinate lawyers, staff, and outside service providers are being supervised. Those are positive obligations. They do not disappear because a model writes in polished prose.

A Recent Court Signal: Referral Can Matter More Than the Fine

The Minnesota Tax Court's May 29, 2025 order in Delano Crossing 2016, LLC v. County of Wright is useful here because it shows that the risk is not limited to an invoice-sized penalty. The court found that five citations in a summary judgment brief were generated by artificial intelligence and did not refer to real cases. The court declined to impose monetary sanctions, but it referred the matter to the Minnesota Lawyers Professional Responsibility Board.

That is an important professional-ethics point. A lawyer can avoid the most dramatic financial sanction and still face discipline-related consequences. For a practitioner or firm, that changes the design problem. The goal is not simply "reduce sanction exposure." The goal is prevent unsupported authority from ever reaching the filing stage.

What a Real AI-Supervision Protocol Looks Like

A defensible protocol has to be boring. Every case citation should be opened in the original source or a trusted database. Every quoted sentence should be matched back to the source text. Every record cite should be verified against the underlying filing or exhibit. And every AI-generated research memo should be treated as draft research, not filing-ready authority.

The same is true for confidentiality and client communication. Formal Opinion 512 does not say lawyers must avoid AI altogether. It says they must evaluate the benefits and risks of the tools they use. That means understanding whether client inputs are retained, whether outside vendors are involved, and whether the client should be informed or asked for consent because AI use affects how the representation is being carried out.

None of this means AI has no role in legal work. It means AI output does not substitute for professional judgment. It can accelerate issue spotting, organize source material, and draft rough language. It cannot satisfy the lawyer's duty of candor, competence, or supervision by itself. The attorney remains the accountable actor.

The Better Product Standard

Legal tools should be designed around that professional reality. A strong tool should push the lawyer back to sources, make uncertainty obvious, and preserve a trail of what was verified. A weak tool does the opposite: it creates fluent text that looks finished before the verification work has happened.

Legal Exception is built for that stricter standard. The point is not to replace attorney judgment. The point is to help lawyers surface exceptions, carveouts, and overlooked authorities while keeping source review and professional responsibility where they belong: with the attorney.